Multiple sources in the same location

We now consider paths from pairs of hosts in the same location but on entirely different networks to destinations in the UnivHosts set. We consider two such pairs of traceroute sources: (a) a machine on the Berkeley campus and another also in Berkeley but on @Home's cable modem network, and (b) a machine at the University of Washington (UW) campus in Seattle and another on the Microsoft Research network 10 km away.

Figure 4: CDF of distance ratio for paths from pairs of co-located sources to UnivHosts.

Figure 4 shows the CDF of the distance ratio for all 4 sources. For the two sources located in Berkeley, we find that the one on the university campus has a significantly smaller distance ratio, especially at the tail of the distribution. For instance, the 90th percentile of the distance ratio for the UC Berkeley source is 1.41 while that for the cable modem source is 1.83. Since the destination set is UnivHosts, the UC Berkeley source tends to have more direct routes (via Internet2) than the cable modem client has (via @Home and other commercial ISPs). We observe a similar trend for the UW-Microsoft pair. The UW source has more direct routes to other university hosts than does the Microsoft source. For instance, the path from Microsoft to the University of Chicago follows a highly circuitous route through BBNPlanet's (Genuity) network. The geographic path traversed includes Los Angeles, Carlton (TX), Indianapolis and Chicago (in that order). The linearized distance of the path is 4976 km while the geographic distance between Seattle and Chicago is only 2795 km. In contrast, the path from UW (via Internet2) is far more direct: it passes through Denver, Kansas City, Indianapolis, and finally Chicago, for a total linearized distance of 3533 km. These results indicate that the nature of network connectivity of the source and the destination has a significant impact on how direct or circuitous the network paths are.